Here is a short description of the problem I am trying to solve: I have test data for multiple variables (weight, thickness, absorption, etc.) that are taken at varying intervals over time - no set schedule, sometimes a test a day, sometimes days might go between tests. I want to detect trends in each of these and alert stake holders when any parameter is trending up/down more than a certain amount. I first did a linear model between each variable's raw data and test time (I converted the test time to days or weeks since a fixed date) and create a table with slopes for each variable - so the stake holders can view one table for all variables and quickly see if any of them is raising concern. The issue was that the data for most variables is very noisy. Someone suggested using time series functions, separating noise and seasonality from the trends, and studying the trend component for a cleaner analysis. I started to look into this and see a couple concerns/questions already:
Time series analysis seems to require specifying a frequency - how do you handle this if your test data is not taken at regular intervals
If one gets over the issue in #1 above, decomposes the data, and gets the trend separated out (ie. take out particularly the random variation/noise), how would you then get a slope metric from that? Namely, if I wanted to then fit a linear model to the trend component of the raw data (after decomposing), what would be the x (independent) variable? Is there a way to connect the trend component of the ts-decompose function with the original data's x-axis data (in this case the actual test date/times, say converted to weeks or days from a fixed date)?
Finally, is there a better way of accomplishing what I explained above? I am only looking for general trends over time - say over 3 months of data, not day to day trends.
Time series are generally used to see if previous observations of a variable have influence on future observations. You would model under the assumption that the previous observations are able to predict the future observations. That is the reason for that most (not all) time series models require evenly spaced instances of training data. If your data is not only very noisy, but also not collected on a regular basis, then you should seriously consider if time series is the appropriate choice of modelling.
Time series analysis seems to require specifying a frequency - how do you handle this if your test data is not taken at regular intervals.
What you can do, is creating an aggregate by increasing the time bucket (shift from daily data to a weekly average for instance) such that every unit of time has an instance of training data. Following your final comment, you could create the average of the observations of the last 3 months of data instead from the observations.
If one gets over the issue in #1 above, decomposes the data, and gets the trend separated out (ie. take out particularly the random variation/noise), how would you then get a slope metric from that? Namely, if I wanted to then fit a linear model to the trend component of the raw data (after decomposing), what would be the x (independent) variable?
In the simplest case of a linear model, the independent variable is the unit of time corresponding to the prediction you are trying to make. However this is not always regarded a time series model.
In the case of an autoregressive model, this would be the previous observation of what you are trying to predict, something similar to y(t) = x(t-1), for instance multiplied by a smoothing factor. I encourage you to read Forecasting: principles and practice which is an excellent book on the matter.
Is there a way to connect the trend component of the ts-decompose function with the original data's x-axis data (in this case the actual test date/times, say converted to weeks or days from a fixed date)?
The function decompose.ts returns a list which includes trend. Trend is a vector of the estimated trend components corresponding to it's respective time value.
Let's create an example time series with linear trend
df <- data.frame(
date = seq(from = as.Date("2021-01-01"), to = as.Date("2021-01-10"), by=1)
)
df$value <- jitter(seq(from = 1, to = nrow(df), by=1))
time_series <- ts(df$value, frequency = 5)
df$trend <- decompose(time_series)$trend
> df
date value trend
1 2021-01-01 0.9170296 NA
2 2021-01-02 1.8899565 NA
3 2021-01-03 3.0816892 2.992256
4 2021-01-04 4.0075589 4.042486
5 2021-01-05 5.0650478 5.046874
6 2021-01-06 6.1681775 6.051641
7 2021-01-07 6.9118942 7.074260
8 2021-01-08 8.1055282 8.041628
9 2021-01-09 9.1206522 NA
10 2021-01-10 9.9018900 NA
As you see, the trend component already is an estimate of the dependent variable at the corresponding time. In decompose the estimate of trend is based on a moving average.
first post here :) Really simple questions in bold. I want to do kriging using PM10 daily data for 8 static stations in Santiago, Chile, from 1997-2012, into 34 centroids which map different counties. I explain what I've done so far with some questions in between. I'm using just 2008 (less missing values) data as first experiment.
DESCRIPTION:
DATA: I have a column with days from 1997 to 2012 and 8 columns with PM10 station data. I imported to R this data, and established time:
time <-as.POSIXlt(data$date) with no error:
" 1997-04-05 UTC 1997-04-12 UTC.... 2012-12-27 UTC"
I imported stations with it's coordinates and built its projections.
coordinates(stations)=~longitud+latitud
proj4string(stations) <- CRS("+proj=longlat + ellps=WGS84")
In order to create the STFDF, I first built the vector PM10 of data with 8 stations, ordered:
PM10<-as.vector(cbind(data $PM10bosque,data $PM10cerrillos,data $PM10cerronavia,data $PM10condes,data $PM10florida,data $PM10independencia,data $PM10parqueoh,data $PM10pudahuel))
PM10<-data.frame(PM10)
DFPM=STFDF(stations, time, PM10)
DFPM<-as(DFPM,"STSDF")
, the last line because I am working with missing data. Then the estimated variogram and its modelling (which I know it's poor) was done with:
varPM10 <- variogramST(PM10~1,data=DFPM,tunit="days",assumeRegular=F,na.omit=T)
sepVgm <- vgmST("separable",space=vgm(0,"Exp", 8, 700),time =vgm(200,"Exp", 15, 700), sill=100)
sepVgm <- fit.StVariogram(varPM10, sepVgm)
Which results in:
Variograms
Then I used KrigeST this way:
gridPM10 <-STF(centroids,time) (centroids defined previously the same way as stations)
krigedPM10<-krigeST(PM10~1, DFPM, newdata=gridPM10,modelList=sepVgm)
The result of ploting one station data and kriged data for that station county's centroid is:
Kriging result for Cerillos county and its station data
which seems as if the estimation ocurrs for time windows by a set of dates.
First question: Does anybody know why this kriging has this shape?
Then I wondered what would happen if I just used distance as predictor so I coded instead:
varPM10 <- variogramST(PM10~1,data=DFPM,tunit="days", tlags=0:0, assumeRegular=F,na.omit=T)
Second question: Is this a reasonable way to try just distnace as predictor? If not, any advise about how to adjust my code do I can do this is very appretiated. Anyway, this is the result:
Variogram with tlag=0:0
using sepVgm <- vgmST("separable",space=vgm(1,"Per", 8, 700),time =vgm(200,"Exp", 15, 700), sill=100)
By the way, how would you guys fit this?
, then the output really surprised me:
Kriging result with tlags=0:0
Third question: Why I am getting this result? I know the variogram modelling is poor, but even if that's true I understand the program should use the station data of the corresponding date so at least it should change in time.
I have many data sets with known outliers (big orders)
data <- matrix(c("08Q1","08Q2","08Q3","08Q4","09Q1","09Q2","09Q3","09Q4","10Q1","10Q2","10Q3","10Q4","11Q1","11Q2","11Q3","11Q4","12Q1","12Q2","12Q3","12Q4","13Q1","13Q2","13Q3","13Q4","14Q1","14Q2","14Q3","14Q4","15Q1", 155782698, 159463653.4, 172741125.6, 204547180, 126049319.8, 138648461.5, 135678842.1, 242568446.1, 177019289.3, 200397120.6, 182516217.1, 306143365.6, 222890269.2, 239062450.2, 229124263.2, 370575384.7, 257757410.5, 256125841.6, 231879306.6, 419580274, 268211059, 276378232.1, 261739468.7, 429127062.8, 254776725.6, 329429882.8, 264012891.6, 496745973.9, 284484362.55),ncol=2,byrow=FALSE)
The top 11 outliers of this specific series are:
outliers <- matrix(c("14Q4","14Q2","12Q1","13Q1","14Q2","11Q1","11Q4","14Q2","13Q4","14Q4","13Q1",20193525.68, 18319234.7, 12896323.62, 12718744.01, 12353002.09, 11936190.13, 11356476.28, 11351192.31, 10101527.85, 9723641.25, 9643214.018),ncol=2,byrow=FALSE)
What methods are there that i can forecast the time series taking these outliers into consideration?
I have already tried replacing the next biggest outlier (so running the data set 10 times replacing the outliers with the next biggest until the 10th data set has all the outliers replaced).
I have also tried simply removing the outliers (so again running the data set 10 times removing an outlier each time until all 10 are removed in the 10th data set)
I just want to point out that removing these big orders does not delete the data point completely as there are other deals that happen in that quarter
My code tests the data through multiple forecasting models (ARIMA weighted on the out sample, ARIMA weighted on the in sample, ARIMA weighted, ARIMA, Additive Holt-winters weighted and Multiplcative Holt-winters weighted) so it needs to be something that can be adapted to these multiple models.
Here are a couple more data sets that i used, i do not have the outliers for these series yet though
data <- matrix(c("08Q1","08Q2","08Q3","08Q4","09Q1","09Q2","09Q3","09Q4","10Q1","10Q2","10Q3","10Q4","11Q1","11Q2","11Q3","11Q4","12Q1","12Q2","12Q3","12Q4","13Q1","13Q2","13Q3","13Q4","14Q1","14Q2","14Q3", 26393.99306, 13820.5037, 23115.82432, 25894.41036, 14926.12574, 15855.8857, 21565.19002, 49373.89675, 27629.10141, 43248.9778, 34231.73851, 83379.26027, 54883.33752, 62863.47728, 47215.92508, 107819.9903, 53239.10602, 71853.5, 59912.7624, 168416.2995, 64565.6211, 94698.38748, 80229.9716, 169205.0023, 70485.55409, 133196.032, 78106.02227), ncol=2,byrow=FALSE)
data <- matrix(c("08Q1","08Q2","08Q3","08Q4","09Q1","09Q2","09Q3","09Q4","10Q1","10Q2","10Q3","10Q4","11Q1","11Q2","11Q3","11Q4","12Q1","12Q2","12Q3","12Q4","13Q1","13Q2","13Q3","13Q4","14Q1","14Q2","14Q3",3311.5124, 3459.15634, 2721.486863, 3286.51708, 3087.234059, 2873.810071, 2803.969394, 4336.4792, 4722.894582, 4382.349583, 3668.105825, 4410.45429, 4249.507839, 3861.148928, 3842.57616, 5223.671347, 5969.066896, 4814.551389, 3907.677816, 4944.283864, 4750.734617, 4440.221993, 3580.866991, 3942.253996, 3409.597269, 3615.729974, 3174.395507),ncol=2,byrow=FALSE)
If this is too complicated then an explanation of how, in R, once outliers are detected using certain commands, the data is dealt with to forecast. e.g smoothing etc and how i can approach that writing a code myself (not using the commands that detect outliers)
Your outliers appear to be seasonal variations with the largest orders appearing in the 4-th quarter. Many of the forecasting models you mentioned include the capability for seasonal adjustments. As an example, the simplest model could have a linear dependence on year with corrections for all seasons. Code would look like:
df <- data.frame(period= c("08Q1","08Q2","08Q3","08Q4","09Q1","09Q2","09Q3","09Q4","10Q1","10Q2","10Q3",
"10Q4","11Q1","11Q2","11Q3","11Q4","12Q1","12Q2","12Q3","12Q4","13Q1","13Q2",
"13Q3","13Q4","14Q1","14Q2","14Q3","14Q4","15Q1"),
order= c(155782698, 159463653.4, 172741125.6, 204547180, 126049319.8, 138648461.5,
135678842.1, 242568446.1, 177019289.3, 200397120.6, 182516217.1, 306143365.6,
222890269.2, 239062450.2, 229124263.2, 370575384.7, 257757410.5, 256125841.6,
231879306.6, 419580274, 268211059, 276378232.1, 261739468.7, 429127062.8, 254776725.6,
329429882.8, 264012891.6, 496745973.9, 42748656.73))
seasonal <- data.frame(year=as.numeric(substr(df$period, 1,2)), qtr=substr(df$period, 3,4), data=df$order)
ord_model <- lm(data ~ year + qtr, data=seasonal)
seasonal <- cbind(seasonal, fitted=ord_model$fitted)
library(reshape2)
library(ggplot2)
plot_fit <- melt(seasonal,id.vars=c("year", "qtr"), variable.name = "Source", value.name="Order" )
ggplot(plot_fit, aes(x=year, y = Order, colour = qtr, shape=Source)) + geom_point(size=3)
which gives the results shown in the chart below:
Models with a seasonal adjustment but nonlinear dependence upon year may give better fits.
You already said you tried different Arima-models, but as mentioned by WaltS, your series don't seem to contain big outliers, but a seasonal-component, which is nicely captured by auto.arima() in the forecast package:
myTs <- ts(as.numeric(data[,2]), start=c(2008, 1), frequency=4)
myArima <- auto.arima(myTs, lambda=0)
myForecast <- forecast(myArima)
plot(myForecast)
where the lambda=0 argument to auto.arima() forces a transformation (or you could take log) of the data by boxcox to take the increasing amplitude of the seasonal-component into account.
The approach you are trying to use to cleanse your data of outliers is not going to be robust enough to identify them. I should add that there is a free outlier package in R called tsoutliers, but it won't do the things I am about to show you....
You have an interesting time series here. The trend changes over time with the upward trend weakening a bit. If you bring in two time trend variables with the first beginning at 1 and another beginning at period 14 and forward you will capture this change. As for seasonality, you can capture the high 4th quarter with a dummy variable. The model is parsimonios as the other 3 quarters are not different from the average plus no need for an AR12, seasonal differencing or 3 seasonal dummies. You can also capture the impact of the last two observations being outliers with two dummy variables. Ignore the 49 above the word trend as that is just the name of the series being modeled.